Application of the Double Metaphone Algorithm to Amharic Orthography

13 Pages·2004·0.38 MB·English

by  Daniel Yacob

Application of the Double Metaphone Algorithm to Amharic Orthography
International Conference of Ethiopian Studies XV
Daniel Yacob

Abstract
The Metaphone algorithm applies the phonetic encoding of orthographic sequences to simplify words prior
to comparison. While Metaphone has been highly successful for the English language, for which it was
designed, it may not be applied directly to Ethiopian languages. The paper details how the principles of
Metaphone can be applied to Ethiopic script and uses Amharic as a case study. Match results improve as
specific considerations are made for Amharic writing practices. Results are shown to improve further when
common errors from Amharic input methods are considered.

Introduction
In the field of text analysis the concept of word distance is introduced and becomes of central importance to
the problem. Word distance is the essence of word comparison which appears in innumerable problems of
text processing. Text searching and document spell checking are but a few familiar examples.
The notion of distance between groups of written symbols may at first seem highly abstract since
we are trained to think of distance as intrinsically spatial. In the natural world we know distance as a
geometric orientation between two tangible objects, measured with an equally tangible instrument. Word
distances are unitless and determined by the number of corrections required, character by character, to
transform one word into another. Each type of correction is given a (likewise unitless) “weight” value
where the magnitude may reflect the severity of the discrepancy. The sum of these values then determines
the overall magnitude of the word distance.
In naive form all symbols will be treated equally. A set of words of equal length with completely
unlike phonemic sequence will appear equally distant from a comparison word. To improve upon the
distance distribution, linguistic and orthographical information may be applied as a stage of the distance
analysis. Most commonly, linguistic analysis is applied to precondition words to better improve their
apparent proximity from the comparison word.
The preconditioning approach is particularly useful and has as its aim the rectification of the
disparity between the phonology of the spoken word and the orthography of the written word. For example
vowel clusters would be represented by single symbol.
This paper presents an investigation into the application to Amharic orthography of a very
successful preconditioning technique, the Metaphone algorithm, used for English and European languages.
Before presenting the details of the technique itself, it is useful to first review the nature of the problem that
it will be applied to.

Problems in Amharic Spelling
Amharic orthography reflects the spoken phonetic features to a large extent. So closely that one can be lead
to believe that there is no notion of “spelling” in Amharic. The rule generally followed is “if a word sounds
right when read aloud then it was rightly written”. Upon closer inspection we quickly realize that Amharic
spelling rules are just very forgiving when compared to the strict, albeit irregular, conventions of English.
When compared to English, the Amharic author is to a degree liberated to put more cognitive energy into
qualitative writing and less into how words must be written. Though many fewer and requiring less
conscious attention, Amharic spelling very clearly has rules. For example, some phonetical spelling
variations are more acceptable than others. While few would give pause to “ውኃ” vs “ዉሀ”, or even
“ታህሳስ” vs “ታኅሣሥ”, the rendering “ዓዲሥ Aበባ Iትዮጵያ” while phonetically valid is not an acceptable
replacement for “Aዲስ Aበባ Iትዮጵያ”. Though we will explore a number of other categories of Amharic
misspellings, spelling error correction may be viewed in large part as a standardization effort.

Levels of Amharic Spelling
In Ethiopian society there are acceptable levels of precision, a phono-orthographical radius that renderings
may fall within to be considered recognizable and acceptable. The sarcastic quip of Mark Twain:

“I respect a man who knows how to spell a word more than one way”
becomes the modus operandi of Amharic spelling. At the most basic level words may be spelt identically as
they are spoken. The basic level provides a working medium for all informal exchanges. The basic level is
the most pragmatic and so long as character sequences can be converted back into the sounds of the words
they record, no judgment is passed over the spelling quality (e.g. “ባል”, “መልክት”,“Eንቁጣጣሽ”, “Aይን”,
“Aሳ”, “ያገር”). The glyph forms of letters may also be found to vary here. Persons with minimal education
up to high school level will not often venture beyond the basic spelling conventions.
At the intermediate level, more formal word choice is found in diction and variations in word
renderings diminish. The writer is likely making a conscious effort to spell with the canonical syllograph
from amongst available isophonic alternatives. (e.g. “በዓል”, “መልEክት”, “Eንቍጣጣሽ”, “ዓይን”, “Aሣ”,
“የAገር”). The middle level of conformance is expected in novels, government newspapers, and professional
literature.
At the most advanced level of writing the canonical forms of words must be used. Words can no
longer be written merely as they would be spoken. This highest, hypercorrect, level of formal Amharic is
largely the realm of specialists. Literature demanding this level of conformance would be Biblical, imperial
and dictionaries. Accordingly, these works also become references for canonical Amharic. Writers at the
top level have almost invariably had training in Amharic language and literature and usually in Ge’ez as
well. (e.g. “በAል”, “መልEክት”, “Eንቍጣጣሽ”, “Aይን”, “Aሣ”, “የሀገር”).
Level boundaries are by no means discrete, rather they are blurred (particularly in the lower levels)
and practices normally found at a higher level may in fact be found mixed within a manuscript that
predominately exhibits features of a lower level. As might be expected, convention drift occurs more in the
downward direction than upward. An Egyptian or Aztec pyramid model is somewhat useful to consider.
What these structures help illustrate are the wider degree of spelling variance considered acceptable at the
base and greater conformance towards the canonical at the peak. To further define the three level model
and debate features such as shape, thickness, and sublevels is of limited practical value for the problem of
word rendering validation. The three levels are not formally recognized and offer no more than a heuristic
model.
Arguments can be made for software to support the spelling conventions of the top two levels.
Amharic spelling software should always support the canonical renderings. Though in a number of cases
correctly rendered words may not be immediately recognized by the average person and will appear out of
place in the informal manuscript. Time will tell. User feedback and perhaps a market survey would be
required to determine if the software support for pencanonical conventions would be worth developing.

Errors Inherent from Symbol Redundancy
As it is well known, much of Amharic’s vocabulary has its origins in the Ge’ez language. Amharic
maintains the entirety of Ge’ez symbology while preserving most, but not all of, Ge’ez’s phonology. As a
consequence, the phonemes of a number of symbols in the orthography have decayed into the nearest
Amharic equivalent. The resulting redundancy in symbols in Amharic is easily the most prevalent and
readily observable influence that complicates Amharic spelling. For the purpose of future reference we will
label this issue as “Type 1: Syllographic Redundancy”.
While the symbols are homophonic equivalents they are not interchangeable at the top level. At
the lower levels, while the homophonic characters may be interchanged freely not all substitutions that are
possible are also probable. Nonetheless, an Amharic spellchecker should be capable of equating all
homophonic equivalents and be able to offer the canonical form.

Table 1: Syllographic Redundancies
Possible but Improbable
Canonical Amharic Common Amharic
Amharic
ዓለም Aለም Aለም, ኣለም
ጸሀይ, ጸሓይ, ጸኀይ, ጸኃይ, ጸኻይ,
ፀሐይ ጸሃይ, ፀሃይ, ጸሐይ
ፀሀይ, ፀሓይ, ፀኀይ, ፀኃይ, ፀኻይ
ኀይለ ሥላሴ7, ኃይለ ስላሴ, ኀይለ ስላሴ,
ኀይለ ስላሤ, ኃይለ ስላሤ, ሀይለ ስላሴ, ሐይለ ሥላሤ, ሓይለ ሥላሤ, ሓይለ
ሀይለ ስላሤ, ሀይለ ሥላሴ, ሃይለ ስላሴ, ሥላሴ, ሓይለ ስላሤ, ሓይለ ስላሴ,
ኃይለ ሥላሴ
ሃይለ ሥላሴ, ሃይለ ስላሤ, ኀይለ ሥላሤ, ኻይለ ሥላሤ, ኻይለ ሥላሴ, ኻይለ
ኃይለ ሥላሤ, ሀይለ ሥላሤ, ሃይለ ሥላሤ, ስላሤ, ኻይለ ስላሴ
ሐይለ ስላሴ, ሐይለ ስላሤ, ሐይለ ሥላሴ
2
In addition to the phonetic redundancy of characters, Amharic suffers slightly from visual redundancy in a
few cases. Most prominently the vowel markers of; ‘ው’ and ‘ዉ’ , ‘ፕ’ and ‘ኘ’, ‘ፖ’ and ‘ፓ’, ‘ዪ’ and ‘ዩ’,
‘ጕ’ and ‘ጒ’ are similar enough that the former characters are often interchanged in words with the later.
This problem is exacerbated at small print sizes and with the lack of visual clarity often found on computer
screens. Additionally, the letter ‘ቁ’ is often used in place of ‘ቍ’ (e.g. “ቍጥር” vs “ቁጥር”) which may owe
more to phonetic proximity and decay than to visual (again type face quality plays a role here). Often times
the writer may simply be choosing the form that is easiest to write by hand or type into a computer.
A similar problem that stems from confusion with Ethiopic vowel markers occurs with the arm-
and-ring vowel marker used for both 7th and 8th form syllables. The arm-and-ring marker makes its first and
most frequent appearance on the 7th syllable ‘ሎ’ for which the sign is then most associated. The arm-and-
ring-marker shows up again with the infrequently used 8th form elements ‘ቈ’, ‘ኈ’, ‘ኰ’ and ‘ጐ’ in the
Ge’ez character set. Given the infrequent and somewhat specialized use of these rounded labial elements,
their purpose is often forgotten by the layman and they become used as alternatives to their 7th form counter
parts. Note that ‘ጐ’ and ‘ኰ’ are the most often substituted letters, and that the rules reverse for ‘ቈ’ which
becomes the substitutee, though substitutions may occur in both directions for all elements.

Table 2: Glypheme Misidentifications
Canonical Amharic Common Amharic
ቈረሰ ቆረሰ
መኰንን መኮንን
ጎንዳር ጐንዳር

It is noteworthy that the arm-and-ring also gets added to ‘የ’ as an alternative rendering of the 7th member of
the ‘የ’ family. Also deserving of attention are letter face design styles whereby ‘ጨ’ and ‘ጬ’ will appear
as ‘î’ and ‘ጨ’ respectively. Occasionally an author who is more familiar with the later style may out of
habit apply ‘ጨ’ in place of ‘ጬ’ when it is inappropriate for the working typeface. We will later refer to
this issue of letter appearance problems as “Type 2: Glypheme Misidentification”.

Errors Inherent from Phonology-Orthography Disconnect
As another consequence of the decline of Ge’ez phonemes in Amharic discussed earlier, Ge’ez words in
Amharic may be rendered identically or may have lost one or more letters over time. In some cases letters
that do remain in canonical Amharic rendering have become silent. Having become superfluous, the silent
letters are commonly dropped or elided at the basic and intermediate levels. A writer working at the
intermediate level may wish to use the canonical rendering but insert the wrong letter. The Ge’ez style
pluralization is another example of a “false Ge’ezism” that may occur when a word is thought to be of
Ge’ez origin or maintain Ge’ez rules. We will label this issue “Type 3: False Ge’ezisms”. In some
instances, mostly in older materials that were not electronically produced, a false Ge’ezism may in fact be
an archaic Amharic spelling:

Table 3: False Ge’ezisms
False Ge’ezism /
Canonical Amharic Ge’ez
Archaic Amharic
ምልክት ምልEክት ምልክት7
ቀለሞች ቀለማት ቀለማት
ትEግሥት ትEግEሥት ትEግሥት
Aየ15 Aየ15 ርEየ
ሁሉ ኍሉ, ኹሉ6 ኵሉ7
ሁለት ኍለት, ኹለት6 ክሌቱ7

There are a number of common cases where phonology clashes with Amharic orthography. For instance,
‘ም’ may be exchanged for ‘ን’ before ‘በ’, as in “Aንበሳ” vs “Aምበሳ”. Likewise ‘ም’ may also replace ‘ን’
before forms in ‘ፈ’ (e.g. “ላንፋ” vs “ላምፋ”).15 Spoken Amharic has a great many alternations, whole and
partial assimilations. Not all spoken occurrences will also manifest themselves in written form. For the sake
of simplicity and efficiency we attempt to account only for occurrences encountered in surveyed literature.
“Type 4: Assimilations and Alternations”.
3
The occurrence of the morpheme –(i)ያ- is a source of much confusion. This is no doubt due to the
similarity between the phonology of collective nouns and the noun of instrument. The nouns of instrument
can be considered a derivative of the infinitive where the suffix “+iያ” is appended causing the ending 6th
form syllable to transform into the 3rd (e.g. “መስበር” ⇒ “መስበሪያ”).15 Some collective nouns work the
opposite way where the 3rd from will transform into the 6th as the noun of agent of agent is pluralized with
-iያን (e.g. “Aንባቢ” ⇒ “Aንባብያን”). “ክርስቲያን”3 is a special case that breaks the collective noun rule:

Table 4: Phonology-Orthography Disconnects in –(i)ያ-
Canonical Amharic Common Amharic
Aንባብያን Aንባቢያን
ወንጌላውያን ወንጌላዊያን
Aብዮታውያን Aብዮታዊያን
Iትዮጵያውያን Iትዮጵያዊያን

Though it has not been stated until now, our focus is on underived words. Any and all affixes must be
stemmed before the Metaphone algorithm can be applied. The complexity with –(i)ያ- will also impact the
rendering of underived nouns that feature a natural 6th form syllable plus ‘ያ’. In such cases the 6th form
syllable may be rendered in the 3rd form:

Table 5: Additional Phonology-Orthography Disconnects in -iያ-
Canonical Amharic Common Amharic
ሚያዝያ ሚያዚያ
ማርያም ማሪያም
ሰማንያ ሰማኒያ

Phoneme-morpheme boundaries are generally preserved in canonical Amharic but may be lost in prefix
elisions in the orthography (e.g. “ያገር” vs “የAገር”, “ያማርኛ” vs “የAማርኛ”, “በንደ” vs “በEንደ”). Elisions
may also occur when an alef-vowel follows a 6th form syllable “መልAኩ” ⇒ “መላኩ”. The elision
occurrences we will label “Type 5: Orthographic Abbreviations and Elisions”.
Labio-velarizations present a problem similar to the -iያ- morpheme. It is not readily predictable
when the ‘ዋ’ inflection should join with the preceding syllable or remain separate (e.g. “Eሷ” vs “Eስዋ”).
The confusion with labio-velar renderings we will label now “Type 6: Disjoint Labiovelars”:

Table 6: Disjoint Labiovelars
Canonical Amharic Common Amharic
ሆኗል ሆኖዋል, ሆኖAል
ጡዋት ጥዋት, ጠዋት, ጧት
ይዟል ይዞAል, ይዞዋል, ይዝዋል

Regional dialects can also impact word formation in the basic level where the words are more likely to be
written following their spoken form; “ሂጂ” vs “ሂጅ”, “Aይዶለም” vs “Aይደለም”, “ዓጤ” vs “ዓፄ”.
An Amharic spellchecker should be able to associate the canonical, common and commonly
misrendered forms together. Regional and dialect variations that impact spelling we will label “Type 7:
Dialect Variations”.

Errors Inherent from Foreign Language Transcription
When words arrive to Amharic from outside languages, occurring very often for technology terms, the
assimilation process requires at least 3-5 years. Until a stable convention emerges, contrasting
transcriptions will be found in popular literature. An Amharic spellchecker should be able to recognize all
candidates as related and later present the favored convention once identified. The conflicting transcriptions
problem we will now label “Type 8: Foreign Language Transcription”:

4
Table 7: Foreign Language Transcriptions
Canonical Amharic Common Amharic
ቴክኖሎጂ3 ቴክኒዎሎጂ16
ኮምፒዩተር3 ኮምፒውተር16
Iሜይል3 Iሜል, ኤሜል, ኤሜይል

Errors Inherent from the Typing System
Every means of lettering, from a stone stylus to a PDA stylus has its own innate advantages and hazards.
Keyboard composition is the most widely used form of rapid personal character rendering in the present
day. Keyboards offer faster character rendering by removing from the writers hand the burden of forming
the shape of each letter. The keyboard offers an interface to a mechanical or electrical apparatus that will
create the shape of each letter more quickly and consistently than a human ever could. The writer need only
tap the symbol of the desired character on a keyboard “key”.
With proficiency, the writer will use all fingers to strike keys independently and without
observation or even thought of what the hands are doing. Word formation errors that are likely from
keyboard composition are related to the geometric distribution of the letter symbols, the skill of the writer,
and the complexity of the mechanical motions required of the hands to compose a letter.
As the number of keys on a keyboard is invariably fewer than the number of symbols of the
Ethiopic writing system; Ethiopic input methods must be defined such that the majority of letters will
require that several keys be struck for composition. The possibility exists in the course of composition that
the writer will omit a key or strike a key other than the intended (a “mistrike”). Left uncorrected these
mistrikes will lead to a word formation error.
The type of mistrikes that may occur depends highly on the input method. Mistrikes that are
common under one input method (such as the Ethiopian typewriter) may be less likely or impossible with
another (such as phonetic based). Independent from the keyboard itself, the rendering apparatus may not
make available a desired letter or symbol. In these cases the writer may elect to substitute the unavailable
letter with another, omit it, or compose a facsimile of the symbol thru other available symbols and
glyphemes. We will apply the common name here, “mistrike”, to identify errors of this class as “Type 9:
Mistrikes”.
To this end we consider the common mistrikes inherent to phonetic based input methods. The
most familiar example mistrikes from phonetic based input methods come from the “shift-slip” condition.
Whereby a typist’s fingers have pressed or released the shift key either too soon or not soon enough
resulting in the wrong letter being composed.

Table 8: Mistrikes of Phonetic Input Methods
Shift-Slip ሐቐኘኸዽጝጠጨጸጰዠ ሀቀነከደግተቸሰጰዘ
5th form for the 1st ጤና ጠና
Unintended Elision AርAያ Aራያ

Metaphone and Sound Based Word Comparison
The Soundex (or Miracode), Metaphone and Double Metaphone algorithms belong to the family usually
known as “phonetic encoding” or “sound alike” algorithms that offer a heuristic type of fuzzy matching.
The algorithms were developed to address the problem of indexing personal names in English which often
come from non-English languages and consequently result in a variety of renderings. The Soundex
algorithm was devised in 1935 to index the US Census results of 1880-1920. The Metaphone algorithm was
introduced more than half a century later in 1990 and offers considerable improvements over its
predecessor. These algorithms take the approach of converting a word into a simplified form based on the
phonetic patterns that the letter sequences represent. The resulting simplified words will then appear
“nearer” to one another in written form.8,13
In essence these name matching algorithms attempt to untangle the rules of English orthography
so that names like “Steven”, “Stephen” and “Stefan” will be recognized as spelling variations of the same
name. Variations on personal name spellings abound in English (e.g. Nelson, Nielson, Niellson, Nelsonn,
Nelsone, Nellson, Nielsone, Nielsen, etc) and while none are deemed “wrong” it is useful to be able to
associate them together. In particular the field of genealogy relies heavily on these algorithms for just this
purpose.
5
Soundex and its decedents are still very much used in the present day. While designed as a tool to
undo the problems inherent with English transcriptions of proper nouns, the algorithms are applied to every
part of speech. With the rise of word processors and desktop publishing the algorithms have a new found
and vital role at the heart of spelling correction software (aka “spellcheckers”). Under this new utilization
the algorithms are applied not to match variations of valid spellings but rather to find a valid word spelling
when presented with one known to be invalid.

Metaphone for Amharic
The Metaphone algorithm, like the other members of its family, exists to serve as tools to operate on
English words and in the English alphabet. Without modifying the Metaphone algorithm, it could not be
applied to Amharic unless Amharic words were first converted into a representation under the English
alphabet. Doing so we would anticipate a lower success rate for Amharic than we find for English as rules
of English orthography are applied out of context on symbol sequences that do not represent English words.
Such an approach would however be useful as a basis of comparison for a Metaphone algorithm derived
specifically for Amharic.
To develop a Metaphone algorithm for Amharic we will want to take into consideration the
problems that we have identified earlier that are particular to Amharic spelling. Firstly, we will assume
throughout that we are working in the Ethiopic character set. Next, we would want our algorithm to be
insensitive to the distinctions between symbols of differing appearance but of like sound. Somewhat
analogous to how the Metaphone algorithm for English ignores letter case by converting words into a
single case, we want to convert phonetically redundant letters into a single representative. Thus we have the
following conversion rules:

Table 9: Simplification of Phonetically Equivalent Syllables
Phonemic Equivalents Simplification
ሀ, ሃ, ሐ, ሓ, ኀ, ኃ, ኻ ሀ
ሑ, ኁ, ኹ ሁ
⋮ ⋮ ⋮ ⋮
ሰ, ሠ ሰ
ሱ, ሡ ሱ
⋮ ⋮ ⋮
A, ኣ, A, ዓ A
U, U U
⋮ ⋮ ⋮
ጸ, ፀ ጸ
ጹ, ፁ ጹ
⋮ ⋮ ⋮
ቁ, ቍ ቍ
ቆ, ቈ ቈ
ኮ, ኰ ኮ
ጎ, ጐ ጎ

An example word “ዓለምፀሐይ” would become “Aለምጸሀይ” under these conversions. In so doing we
overcome the Type 1 problem. The next step that we will perform will be to apply the Metaphone rule of
removing all vowels from a word except an initial vowel should a word start with a vowel. The elimination
of vowels is the most heavy-handed step in the Metaphone process, but it is in large part the key to its
success. Eliminating the vowels in Amharic words takes care of the bulk of the matching issues identified
earlier. Vowels are eliminated by removing all Alef-A (A) and Aynu-A (A) forms, and then by converting
all syllables into the Sadis, or 6th, form. Applied to our example word, “Aለምጸሀይ” simplifies further to
“Eልምጽህይ”. In this step we overcome the identified problems of Type 2, 4 and 5 as well as some
occurrences of Types 7, 8 and 9.
In this simplification process we do make two special exceptions. The first is to preserve ‘ኘ’ in its
Ge’ez form which we will want to compare to ‘ፕ’ shortly to address the glyph mismatch problem. The
6
other exception is to convert the labiovelar forms of the fourth order into two Sadis letters. The first would
be the Sadis form of the syllable itself, followed then by a Sadis ‘ው’. Doing so allows us to compare
renderings for the Type 6 disjoint labiovelar problem. For example; ‘ኳ’ reduces to “ክው” and the
remaining labializations of ‘ከ’ (‘ኰ’, ‘ኵ’, ‘ኲ’ and ‘ኴ’) ultimately become ‘ክ’ as no examples of Type 6
problem have been found for these forms. We continue on to handle spelling mistakes that arise from glyph
confusion, common typing errors, and phonology issues.
Under the Double Metaphone algorithm, a follow up to the Metaphone algorithm arriving in the
year 2000, two simplified phonetic encodings of a word can be generated. The first encoded word is
considered the most probable as a representation for the target word. When appropriate, a second encoded
word is created that offers an alternative encoding under less frequently used English conventions. For
Amharic this approach remains useful and we will apply it as alluded to earlier. However, for the purposes
of investigation we do not want to restrict the possible outcomes to only two encodings. Rather, we will
develop as many encodings as appear to be useful under Amharic writing conventions.
The letter ‘ኘ’ presents a case where we would want to encode the letter canonically as ‘ኝ’ and
alternatively as ‘ፕ’ for cases where the letter identities were mistaken (we only need do this for the first
form of ‘ኘ’ as the mistaken identity problem does not occur with the other syllabic forms). Similarly we
will require an additional encoding of ‘ዽ’ for ‘ጵ’ as part of our treatment for the Type 2 problem.
Unfortunately these measures do not go far enough and ‘ኘ’ and ‘ዸ’ could have both been results of the
shift-slip condition of Type 9 problems. Yet another set of encodings will be created for this possibility
which requires an encoding of ‘ድ’ for ‘ዽ’ and ‘ን’ for ‘ኝ’. The full set of shift-slip substitutions are in the
following table, note that the substitutions occur in both directions (uppercase-for-lowercase and
lowercase-for-uppercase):

Table 10: Mistrikes of Phonetic Input Method
Lowercase Uppercase
ስ ጽ
ቅ ቕ
ት ጥ
ች ጭ
ን ኝ
ክ ኽ
ዝ ዥ
ግ ጝ
ፕ ጵ

Operating on our example word for case mismatch issues we would create the second encoded
word “Eልምስህይ” which we will carry along with “Eልምጽህይ” for all future operations. The step of
checking for case mismatches, which addresses problem Type 9, is worthy of further discussion. It must
first be noted that doing so is closely tied to an assumed input method. The table here presents shift variants
that are applicable to many popular phonetically based input methods in use presently. However, the table
is totally inappropriate for analyzing text that would have been keyed using an input method based on the
Ethiopian typewriter for example. An Ethiopian typewriter based input method has a completely different
set of error types associated with it and would require a different corrective procedure at this stage. The
“keyboard distance” between words should always follow a Metaphone procedure. However, given the
special and high frequency nature of the issue for Ethiopic input methods, the shift-slip condition is
analyzed at this stage. Accordingly, this step should be used with care and when the input method
associated with the document is unknown, should not be used at all. The alternative encodings from this
step are given the lowest priority.
The Metaphone algorithms for English will encode some phonetically similar consonants together.
Such as the labials ‘b’ and ‘p’ both encode as ‘b’, the dentals ‘t’ and ‘d’ both become ‘t’.8 This encoding
strategy may be advantageous for Amharic as well. Until a much larger and more thorough investigation
can be conducted to determine the merits of doing so, we cautiously encode Amharic consonants
separately. A single exception is made at this time for the encoding of ‘ቭ’ as ‘ብ’ which should not
significantly impact our smaller exploratory investigation as very few Amharic words will use the ‘ቨ’
family, and then only for transcription of foreign words.
7
Finally, the only remaining phonological issue we have to address at this time is the substitution of
‘ም’ for ‘ን’ before ‘ብ’ and ‘ፋ’. Encoded words are created for each.
To help illustrate the algorithm steps we have just discussed in prose, the following table presents
encoding examples for misspellings of “ዓለምፀሐይ”, “ጡዋት”, “ወንበር”, “ፕሬዚደንት” and the correctly
spelled “ላም”:

Table 11: Steps in Amharic Metaphone
Metaphone Step Aለምጸሐይ ጧት ወምበር ኘሬዚዳንት ላም
1) Simplify Aለምጸሀይ ጥዋት ወምበር ኘሬዚዳንት ላም
2) Remove Vowels Eልምጽህይ ጥውት ውምብር ኘርዝድንት ልም
3) Check for ውምብር
Eልምጽህይ ጥውት ኘርዝድንት ልም
Phonological Problems ውንብር
4) Check for ውምብር ኝርዝድንት
Eልምጽህይ ጥውት ልም
Glyph Mismatches ውንብር ፕርዝድንት
5) Check for Input Eልምጽህይ ጥውት ውምብር ኝርዝድንት ኝርዝዽንት ልም
Method Problems Eልምስህይ ጥውጥ ውንብር ፕርዝድንት ፕርዝዽንት
ትውጥ ንርዝድንት ንርዝዽንት
ትውት ኝርዥድንት ኝርዥዽንት
ንርዥድንት ንርዥዽንት
ፕርዥድንት ፕርዥዽንት
ኝርዝድኝት ኝርዝዽኝት
ንርዝድኝት ንርዝዽኝት
ፕርዝድኝት ፕርዝዽኝት
ኝርዥድኝት ኝርዥዽኝት
ንርዥድኝት ንርዥዽኝት
ፕርዥድኝት ፕርዥዽኝት
ኝርዝድንጥ ኝርዝዽንጥ
ንርዝድንጥ ንርዝዽንጥ
ፕርዝድንጥ ፕርዝዽንጥ
ኝርዥድንጥ ኝርዥዽንጥ
ንርዥድንጥ ንርዥዽንጥ
ፕርዥድንጥ ፕርዥዽንጥ
ኝርዝድኝጥ ኝርዝዽኝጥ
ንርዝድኝጥ ንርዝዽኝጥ
ፕርዝድኝጥ ፕርዝዽኝጥ
ኝርዥድኝጥ ኝርዥዽኝጥ
ንርዥድኝጥ ንርዥዽኝጥ
ፕርዥድኝጥ ፕርዥዽኝጥ

The example with the Amharic transcription of “ፕሬዚዳንት” helps illustrate a problem in encoded word
generation that occurs in the fifth step for input method correction. Most plainly the approach leads to the
possibility of a high number of encodings being generated. The occurrence of more than one input method
related error per word is not frequent. The encodings that are generated in step five represent all possible
errors, very few of which are probable and worth the computational cost of further investigation.
We can reduce the number of encodings generated by restricting only one input method error to
appear per encoding. This would give us three encodings total for “ጧት” and eleven encodings for
“ኘሬዚዳንት” which is still too high. The number of input method error encodings can be reduced even
further by using a single “lowest common denominator” encoding that itself represents an encoding that is
a reducible form of all other input method error encodings. With our example input method system, such an
encoding presents itself when we apply the shift-slip condition to each key and in the uppercase to
lowercase direction only. In Step 5 we would then generate:
8
Table 12: Amharic Metaphone Adjusted Step 5
Metaphone Step Aለምጸሐይ ጧት ወምበር ኘሬዚዳንት ላም
5) Check for Input Eልምጽህይ ጥውት ውምብር ኝርዝድንት ልም
Method Problems Eልምስህይ ትውት ውንብር ፕርዝድንት
ንርዝድንት

The properly spelled word “ላም” has been used as a simple example to help demonstrate and further
explain how the matching process works. “ላም” is an easy choice to work with as it is only two characters
in length and was not subject to additional encodings as created in steps 3-5. “ልም” is the only possible
encoding for “ላም”. “ልም” however is a possible encoding for some 48 additional letter sequences. The 49
sequences would come from the first seven forms of “ለ” permuted against the first seven forms of “መ”.
“ላም” is then one of these 49 permutations that we know to be a valid word. The other 48 permutations
would be invalid renderings of “ላም” specifically but do not necessarily represent invalid words in and of
themselves. “ሎሚ” for example is a valid word in the remaining set of 48 though not a valid rendering of
“ላም”:

Table 13: Valid Alternatives for “ሊም”
Valid Matches for Valid Matches for
Invalid Word Error Encoding Error Encoding Error Encoding
(With Inflections) (No Inflections)
ሊም ልም ላም, ላሙ, ላሜ ላም
ለም, ለሙ, ለማ ለም, ለማ
ልም, ልሙ, ልሚ, ልማ ልም, ልሚ, ልማ
ሎሚ ሎሚ

It is useful to reflect at this point on the role that the Metaphone algorithm plays in the process of spelling
correction. When a word, such as “ሊም” is not found in an electronic “dictionary” we have identified it as
an unknown word or an incorrect spelling of a known word. The Metaphone algorithm is then used to
simplify the error word for comparison against the simplifications of known words from the dictionary. The
simplification of “ላም”, “ሎሚ” and a few others will be found to “match” (have zero distance from) the
simplification of the error “ሊም”. These matching words become candidates for the correction. Typically
the list of candidates will then be presented to the typist for selection and substitution for the error word
(“ሊም”). The candidates will be presented in an order of greatest probability for the intended word. The
proximity distance from the error word (is “ላም” or “ሎሚ” closer to “ሊም”?) is a major factor in
determining the probability.
With this understanding, Table 14 presents matches that the algorithm will find for two of our
example words that have numerous common misrenderings.

Table 14: Matchable Likely Misspellings of “ዓለምፀሐይ” and “ጡዋት”
Matchable Misspelling of ዓለምፀሐይ Matchable Misspelling of ጡዋት
(All Variations Encode to Eልምጽህይ) (All Variations Encode to ጥውት)
ዓለምፀሃይ, ዓለምፀሀይ,ዓለምጸሐይ, ዓለምጸሃይ, ዓለምጸሀይ, ጥዋት, ጠዋት, ጧት
Aለምፀሐይ, Aለምፀሃይ, Aለምፀሀይ, Aለምጸሐይ, Aለምጸሃይ, Aለምጸሀይ,
Aለምፀሐይ, Aለምፀሃይ, Aለምፀሀይ, Aለምጸሐይ, Aለምጸሃይ, Aለምጸሀይ

Experimental Results
A test of the algorithm procedures described here was conduction on 116 canonically spelled words with
166 corresponding misspellings. The test sample represents the nine categories of misspelling presented in
this paper. The Amharic Metaphone algorithm was compared against the Double Metaphone algorithm for
the same test set.
The collection of sample words was first transliterated into an English representation so that the
Double Metaphone algorithm could then be applied. The semantics of a transliteration system could have a
significant impact on the success of Double Metaphone. For instance, the degree to which a transliteration
system relied on non-alphabetic elements (punctuation, numerals) could make it more or less amenable to
9
the algorithm. To counter this effect, the Double Metaphone algorithm test was then conducted on a total
of four different transliteration systems:

Table 15: Word Matching Results for 166 Misspelled Forms of 116 Words
Procedure Matches Percentage
Ethiopic Script with Amharic Metaphone2 150/166 90%
SERA Transliteration with Double Metaphone9 105/166 63%
Ethiop Transliteration with Double Metaphone9 99/166 60%
ISO/TC46/SC2 Transliteration with Double Metaphone 9 97/166 58%
Mainz Transliteration with Double Metaphone9 88/166 53%

The results are clearly most favorable to the Amharic Metaphone algorithm where words were matched
90% of the time. The discrepancy between the Ethiop, ISO, and SERA transliteration systems is negligible
and might fluctuate with the size of the test set. At comparable levels these transliteration systems rely on
punctuation to denote letter symbols. While the Mainz system avoids the use of punctuation for letter
symbols altogether, it was likely more susceptible to the English orthography rules applied by Double
Metaphone and consequentially found the lowest score.
The Amharic Metaphone results were later improved by applying the standard Metaphone rule of
treating ‘w’ and ‘y’ under vowels rules. With the analogous letters ‘ወ’ and ‘የ’ treated likewise as vowels
an additional 10 words were matched with their misspellings from the test set bringing the Amharic
Metaphone success rate up to 96%.
While the test set was representative of our nine misspelling categories not all misspellings fell
under the scope of Metaphone’s applicability. A perfect matching score was not anticipated but conducting
the test helps to illuminate the limitations of the approach that follow on distance calculation algorithms
will have to address. In particular the Amharic Metaphone algorithm had difficulty matching Type 7 words
when the consonant component of a syllable would change.

Conclusion & Further Research
It is evident from the limited test conducted that higher matching rates are found for Amharic words when
the Metaphone algorithm is adjusted to apply the rules of Amharic orthography in lieu of those of English.
This positive result has held up as the test set of words has grown in number, a trend that is expected to
continue. However, a much more thorough test with a larger corpus of materials should certainly be
conducted. The development here has been no more than exploratory in nature.
Further research should investigate the benefits of; applying the vowel rules to the ‘ወ’ and ‘የ’
families, encoding the ‘ጠ’, ‘ጸ’ and ‘ፀ’ families together, encoding ‘ፐ’ and ‘ፈ’ with ‘በ’ and possibly ‘ጀ’
with ‘ዠ’11. Additional alternations in writing should be investigated as well as mistrikes that occur from
well established input methods. With a large corpus of materials an over matching (false-positives)
analysis should be made to evaluate the merits of these suggestions and to further optimize the algorithm.
Ultimately any approach to Amharic spelling correction is limited by the reliability of the
reference used for canonical formations. The establishment of a comprehensive and authoritative lexicon
for written Amharic would be the single most valuable resource towards the realization of this eventual
goal.
10